OpenAI’s personal beef up bot has no thought how ChatGPT works

Observe ZDNET: Add us as a preferred source on Google.

ZDNET’s key takeaways

• The ChatGPT app’s beef up bot hallucinates this system’s options.
• OpenAI recently has no in-app trojan horse reporting, a significant omission.
• Contemporary chats display the beef up bot has gotten extra a professional.

One of the crucial early successes corporations have cited in enforcing synthetic intelligence within the undertaking is in customer-facing duties akin to visitor beef up. 

If OpenAI’s personal automatic visitor beef up is the long run, consumers are in for a coarse time. 

Additionally: How to use ChatGPT: A beginner’s guide to the most popular AI chatbot

OpenAI’s automatic beef up serve as within its ChatGPT app shows a stunning lack of awareness of the options of the applying, one thing I discovered after I went in search of lend a hand with a trojan horse within the app.

(Disclosure: Ziff Davis, ZDNET’s father or mother corporate, filed an April 2025 lawsuit in opposition to OpenAI, alleging it infringed Ziff Davis copyrights in coaching and working its AI techniques.)

My quest started with a trojan horse within the iPad Professional model of the ChatGPT app. I used to be the use of a subscription model of ChatGPT, the Plus model, for which I pay $21.78 a month. After a number of months of the use of the app with out a drawback, I all at once spotted that resizing the window of the app led to it to freeze. 

A small trojan horse and a nightmare beef up chat

On an 11-inch or 13-inch iPad Professional, operating any model of iOS that may multitask, you’ll be able to put the ChatGPT app into one in every of 4 conceivable window sizes, from an excessively slim window to 1 somewhat higher this is very similar to the telephone model of the app, to 2 higher sizes, together with a full-screen window.

I discovered that resizing the window to both full-screen or the scale slightly below that, the place just a little little bit of the House Display screen is visual, led to the app to grow to be totally unresponsive. No contact purposes generated any reaction, or even the interface components of the app — textual content, borders, instructed box — had been stretched as though the app may just no longer refresh them to deal with the trade in measurement of the window.

I used to be ready to reliably recreate the issue every time I resized the app, and I used to be ready to reliably replica the issue on a 2nd iPad Professional.

I went to the lend a hand phase of the app, known as Lend a hand Middle, which you’ll be able to to find underneath your account knowledge within the decrease left nook of the sidebar of the app. Not one of the lend a hand articles had been any lend a hand, so I made up our minds to check out the chat serve as that pops up in Lend a hand Middle.

Additionally: A minority of businesses have won big with AI. What are they doing right?

The automatic bot, powered via ChatGPT, steered a number of steps that did not lend a hand, akin to forcing the app to hand over and reopening it, and deleting the app and putting in it once more from the Apple App Retailer.

And, then, the beef up bot all at once started to hallucinate one thing that does not exist within the app. 

“Ship Comments within the App,” was once its advice. “You’ll be able to record this trojan horse immediately from inside the ChatGPT app (in most cases underneath Account or Give a boost to > Record a Downside/Ship Comments). This is helping the builders get logs and would possibly accelerate a repair.”

“Most often,” possibly, however no longer on this case. There is not any a part of the Lend a hand Middle this is “Record an issue” or “Ship Comments.” It does not exist.

There’s a serve as known as “Record,” which may also be decided on throughout the present chat, from the menu within the upper-right nook, that’s not for trojan horse experiences. That serve as is for reporting problems with the present chat, akin to, “Violence and self-harm,” or “criminality.”

Additionally: How people actually use ChatGPT vs Claude – and what the differences tell us

I pointed this out to the bot, which, within the agreeable tone for which ChatGPT is known, conceded it was once no longer in point of fact a way of trojan horse reporting. “Thanks for pointing that out — you are right kind. The ChatGPT iPad app recently simplest lets in reporting particular person messages or chats.” 

(By way of the best way, you’ll be able to see the historical past of your beef up chats from the chat window in Lend a hand Middle if you happen to log in once more from the Lend a hand Middle. Perfect to log in earlier than you interact in any beef up chats, as a way to recuperate a report of the beef up chat later.)

Computerized electronic mail beef up was once simply as dangerous

I then emailed the beef up serve as from the OpenAI developer forum, beef up@openai.com, and described my drawback once more. 

In a sequence of automatic emails, after describing the problem, the beef up serve as repeated the similar recommendation because the in-app beef up bot, together with telling me to “record this trojan horse immediately from inside the ChatGPT app (in most cases underneath Account or Give a boost to > Record a Downside / Ship Comments),” including,  “This is helping the builders get logs and would possibly accelerate a repair.”

Once I protested, the automatic electronic mail once more conceded the mistake, telling me, “Thanks for pointing that out — you are right kind.”

I contacted OpenAI — a human spokesperson — and gained affirmation by the use of electronic mail that there’s no trojan horse reporting serve as:

We do not recently be offering a devoted consumer-facing trojan horse reporting portal for the ChatGPT app. Other folks can to find troubleshooting steps and techniques to proportion comments thru our reliable Lend a hand Middle at help.openai.com.
For security-related problems, we have now a proper trojan horse bounty program thru Bugcrowd: https://bugcrowd.com/engagements/openai. You’ll be able to additionally give comments immediately within the product — for instance, via giving a thumbs down on responses that don’t seem to be useful or via reporting conversations that do not meet your expectancies. We are running to make the automatic beef up replies extra correct and efficient. (There’s a “trojan horse bounty” for builders, however that isn’t the similar as trojan horse reporting or function requests via reasonable customers.)

The loss of trojan horse reporting is sudden, as it is a usual providing with maximum tool. 

Additionally: Even the best AI agents are thwarted by this protocol – what can be done

The extra fun truth is that OpenAI’s bot, possibly educated at the corporate’s huge corpus of construction paperwork, did not have the proper details concerning the app’s options.

If that is the state of shopper lend a hand by the use of AI, it is not simplest no longer an answer, but it surely makes the consumer’s frustration even worse.

A small glimmer of hope?

In my exams, the unique trojan horse persists with the newest app unlock, model 1.2025.280. Via experimentation, I discovered that no longer maximizing the app window is the one resolution. 

Additionally: I built a business plan with ChatGPT and it turned into a cautionary tale

Curiously, in my follow-up exams, the app’s beef up bot not hallucinates that I will record a trojan horse. As a substitute, it states up-front that there’s no technique to record a trojan horse within the app and asks for an outline of the issue that may be forwarded to the builders. The bot has in the end discovered one thing concerning the app it is dwelling within.

No less than that is growth, I guess. 

Get the morning’s most sensible tales for your inbox every day with our Tech Today newsletter.